AI in education — learning to direct, not just to use

You cannot delegate to a machine what you cannot articulate. That constraint is the bottleneck in working with AI — and it is also, conveniently, a description of what education has always been for: making implicit expertise explicit.

Developers got good at AI first, not because programming is special, but because code demands that you say exactly what you want in enough detail that a machine can act on it. The habit transferred. Most people who work with knowledge for a living have never been trained that way.

The gap we are working in

Roughly 95% of students use AI in their coursework. A large share of programs have effective bans on using it for any major assignment. Those two facts describe an educational system that has decided to have no position at all — and the resulting silence is itself a curriculum, teaching students that the tool they will use for the rest of their careers is something to hide.

We think the doom-and-gloom story is oversold, and we would rather be part of the alternative. The lab treats the School of Design & Science as a place to prototype curricular development for an AI-integrated program — built to be studied, and built so others can copy what works.

The longitudinal study

We are running a multi-year study of AI-skill development across two very different courses: APS-I, a small antidisciplinary studio, and a large online Web3/AI course spanning undergraduates, graduate students, and working professionals. Session logs, individual-differences measures, prompt snapshots taken at intervals across the term, and pre/post surveys accumulate for the year-1 pilot cohort.

The methodological commitment that shapes everything else: interaction patterns are data requiring interpretation, not outcomes in themselves. More AI use does not imply better performance, and less does not imply worse. A student who reaches a strong outcome in few exchanges may have excellent specification skills; a student who iterates extensively may be doing the harder work of exploring alternatives and stress-testing what comes back. Treating token counts as a learning signal would be the easy mistake, and the wrong one.

What the studio is actually teaching

AI as a practice, not a subject. Most graduate programs still teach AI as something to read about. Students here leave with explicit, transferable rules for working with it: state the intended behavior before you look at the code; test at the extremes; doubt comments and function names, because the tool will believe them over the code itself; decide what needs confirming and delegate only the confirmation legwork.

Concept depth comes through AI dialogue, not despite it. The moment a formula clicks — a decay rate understood as stubbornness, a parameter understood as the knob it actually turns rather than a symbol — repeatedly arrives in conversation. The AI does not replace the learning. It is the medium of it, and it will explain the same idea five ways without getting impatient.

An honest model of the limits, not an AI cheerleader. A program that only shows AI as a great collaborator is marketing. The interesting result is that students converge on the harder edge too: where AI gets in the way, and what verification work the human still owes the system. Verification can easily cost more than writing the thing yourself — knowing when that is true is part of the literacy, not an objection to it.

It works for students who do not yet code. A self-described non-mathematician can leave able to articulate Bayes’ rule and debug a deliberately broken epsilon-greedy agent. That is not a lowered bar; it is what happens when the barrier was notation rather than the idea.

The collaboration deepens as the work does. Early on, students describe AI as a translator. Later, as a fellow code-reader. Later still, as a Socratic questioner that is most useful when it asks the follow-up question instead of filling in the answer — and as a sparring partner for systems design. Same students, same tool, three working relationships, each suited to a different kind of problem. That progression is the thing we are trying to measure.

Where this goes

This year is a pilot. The point is to get the design right for the cohorts that follow, in conversation with learning scientists who do this for a living — and to keep the work in the open, since a prototype nobody can copy is not a prototype.

Related: Make It So Camp takes the same premise to practitioners outside the university, and the probability tutorial is what happens when this way of thinking about explanation gets applied to a textbook.